Verification-Aware Training fixes speculative decoding's mismatch between how drafts are trained and how they are accepted

arXiv 2608.30135·low signalXBlueskyLinkedInCopy link

arXiv 2608.30135 (submitted 2026-08-31) notes that speculative verification is sequential and discards everything from the first rejection onward, yet draft models are trained by token-level imitation with a fixed per-position weighting that reflects neither fact. VAT adds a lightweight jointly trained binary verification head that supervises whether each position survives verification, plus verification-adaptive weighting that keeps full weight up to each sample's first rejection and re-anchors the decay there. It only changes the training objective, so it layers on top of existing speculative-decoding methods without touching inference.